In some cases small changes in hot loops can show big differences.
But it's difficult to identify these differences.
perf diff currently can only diff symbols (functions). We can also expand
it to diff cycles of individual programs blocks as reported by timed LBR.
This would allow to identify changes in specific code accurately.
With this patch set, for example,
$ perf record -b ./div
$ perf record -b ./div
$ perf diff -c cycles
# Event 'cycles'
#
# Baseline [Program Block Range] Cycles
Diff Shared Object Symbol
# ........
......................................................................
................ ..................................
#
48.75% [div.c:42 -> div.c:45]
147 div [.] main
48.75% [div.c:31 -> div.c:40]
4 div [.] main
48.75% [div.c:40 -> div.c:40]
0 div [.] main
48.75% [div.c:42 -> div.c:42]
0 div [.] main
48.75% [div.c:42 -> div.c:44]
0 div [.] main
19.02% [random_r.c:357 -> random_r.c:360]
0 libc-2.23.so [.] __random_r
19.02% [random_r.c:357 -> random_r.c:373]
0 libc-2.23.so [.] __random_r
19.02% [random_r.c:357 -> random_r.c:376]
0 libc-2.23.so [.] __random_r
19.02% [random_r.c:357 -> random_r.c:380]
0 libc-2.23.so [.] __random_r
19.02% [random_r.c:357 -> random_r.c:392]
0 libc-2.23.so [.] __random_r
16.17% [random.c:288 -> random.c:291]
0 libc-2.23.so [.] __random
16.17% [random.c:288 -> random.c:291]
0 libc-2.23.so [.] __random
16.17% [random.c:288 -> random.c:295]
0 libc-2.23.so [.] __random
16.17% [random.c:288 -> random.c:297]
0 libc-2.23.so [.] __random
16.17% [random.c:291 -> random.c:291]
0 libc-2.23.so [.] __random
16.17% [random.c:293 -> random.c:293]
0 libc-2.23.so [.] __random
8.21% [div.c:22 -> div.c:22]
148 div [.] compute_flag
8.21% [div.c:22 -> div.c:25]
0 div [.] compute_flag
8.21% [div.c:27 -> div.c:28]
0 div [.] compute_flag
5.52% [rand.c:26 -> rand.c:27]
0 libc-2.23.so [.] rand
5.52% [rand.c:26 -> rand.c:28]
0 libc-2.23.so [.] rand
2.27% [rand@plt+0 -> rand@plt+0]
0 div [.] rand@plt
0.01% [entry_64.S:694 -> entry_64.S:694]
16 [kernel.vmlinux] [k] native_irq_return_iret
0.00% [fair.c:7676 -> fair.c:7665]
162 [kernel.vmlinux] [k] update_blocked_averages
'[Program Block Range]' indicates the range of program basic block
(start -> end). If we can find the source line it prints the source
line otherwise it prints the symbol+offset instead.
v5:
---
Only the patch 'perf diff: Use hists to manage basic blocks per symbol'
is changed in v5. Since we still carry block_info in 'struct hist_entry'
so we don't need our own new/free ops for hist entries. And the block_info
is released in hist_entry__delete.
v4:
---
Use source lines or symbol+offset to indicate the basic block.
Changed patches:
perf diff: Print the basic block cycles diff
perf diff: Documentation -c cycles option
v3:
---
In v3, the major change is to move most of block stuffs from
'struct hist_entry' to new structure 'struct block_hist' and
update the code accordingly. But we still have to keep the
block_info in 'struct hist_entry' since we need to compare by
block info when inserting new entry to hists.
Others are minor changes, such as abs() -> labs(), removing
duplicated ops and etc.
Changed patches:
perf diff: Use hists to manage basic blocks per symbol
perf diff: Link same basic blocks among different data
perf diff: Print the basic block cycles diff
v2:
---
Keep standard perf diff format.
Following is the v1 output.
# perf diff --basic-block
# Cycles diff Basic block (start:end)
# ........... .......................
#
-208 hrtimer_interrupt (30b9e0:30ba42)
-157 update_blocked_averages (2bf9f2:2bfa63)
-126 interrupt_entry (c00880:c0093a)
-86 hrtimer_interrupt (30bb29:30bb32)
-74 hrtimer_interrupt (30ba65:30bac4)
-56 update_blocked_averages (2bf980:2bf9d3)
48 update_blocked_averages (2bf934:2bf942)
-35 native_write_msr (267900:26790b)
26 native_irq_return_iret (c00a27:c00a27)
22 rcu_check_callbacks (2febb6:2febdc)
-21 __hrtimer_run_queues (30b220:30b2a3)
19 pvclock_gtod_notify (14ba0:14c1b)
-18 task_tick_fair (2c5d29:2c5d41)
Jin Yao (7):
perf util: Create block_info structure
perf util: Add block_info in hist_entry
perf diff: Check if all data files with branch stacks
perf diff: Use hists to manage basic blocks per symbol
perf diff: Link same basic blocks among different data
perf diff: Print the basic block cycles diff
perf diff: Documentation -c cycles option
tools/perf/Documentation/perf-diff.txt | 17 +-
tools/perf/builtin-diff.c | 385 ++++++++++++++++++++++++++++++++-
tools/perf/ui/stdio/hist.c | 27 +++
tools/perf/util/hist.c | 43 +++-
tools/perf/util/hist.h | 9 +
tools/perf/util/sort.h | 13 ++
tools/perf/util/srcline.c | 4 +-
tools/perf/util/symbol.c | 22 ++
tools/perf/util/symbol.h | 23 ++
tools/perf/util/symbol_conf.h | 4 +-
10 files changed, 535 insertions(+), 12 deletions(-)
--
2.7.4
